rnd = rnd * (1 - wall.clamp(max=1))
rnd = torch.rand(nb, height, width)
+ rnd[:, 0, 0] = 0 # Do not put coin at the agent's starting
+ # position
coins = torch.zeros(nb, T, height, width, dtype=torch.int64)
rnd = rnd * (1 - wall.clamp(max=1))
for k in range(nb_coins):